Skip to contentSkip to conversation

tcp/dccp: avoid parity split for socket-local bind range

PATCH In review [email protected] 2 participants 5 messages Jun 26 – Jun 30

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 tcp/dccp: avoid parity split for socket-local bind range

From: luoxuanqiang <[email protected]>

IP_LOCAL_PORT_RANGE lets applications override the netns ephemeral port range on a per-socket basis. __inet_hash_connect() already treats such a range as an explicit application partition and scans it with step 1 [1].

Do the same in inet_csk_find_open_port(): when a socket-local range is set, walk the whole selected range instead of first splitting it by parity. Keep the existing step-2 parity behavior for sockets using the netns range, so the default bind/connect separation remains unchanged.

[1] https://lore.kernel.org/r/[email protected]

---
 net/ipv4/inet_connection_sock.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
net/ipv4/inet_connection_sock.c +13 -7
index 56902bba54838..ad8af70c92ca3 100644
@@ -323,13 +323,16 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 	struct inet_bind2_bucket *tb2;
 	struct inet_bind_bucket *tb;
 	u32 remaining, offset;
+	bool local_ports;
 	bool relax = false;
+	int step;
 
 	l3mdev = inet_sk_bound_l3mdev(sk);
 ports_exhausted:
 	attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 other_half_scan:
-	inet_sk_get_local_port_range(sk, &low, &high);
+	local_ports = inet_sk_get_local_port_range(sk, &low, &high);
+	step = local_ports ? 1 : 2;
 	high++; /* [32768, 60999] -> [32768, 61000[ */
 	if (high - low < 4)
 		attempt_half = 0;
@@ -342,18 +345,19 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 			low = half;
 	}
 	remaining = high - low;
-	if (likely(remaining > 1))
+	if (!local_ports && remaining > 1)
 		remaining &= ~1U;
 
 	offset = get_random_u32_below(remaining);
 	/* __inet_hash_connect() favors ports having @low parity
 	 * We do the opposite to not pollute connect() users.
 	 */
-	offset |= 1U;
+	if (!local_ports)
+		offset |= 1U;
 
 other_parity_scan:
 	port = low + offset;
-	for (i = 0; i < remaining; i += 2, port += 2) {
+	for (i = 0; i < remaining; i += step, port += step) {
 		if (unlikely(port >= high))
 			port -= remaining;
 		if (inet_is_local_reserved_port(net, port))
@@ -384,9 +388,11 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
 		cond_resched();
 	}
 
-	offset--;
-	if (!(offset & 1))
-		goto other_parity_scan;
+	if (!local_ports) {
+		offset--;
+		if (!(offset & 1))
+			goto other_parity_scan;
+	}
 
 	if (attempt_half == 1) {
 		/* OK we now try the upper half of the range */
Signature · 1 line
2.43.0
#1 lore raw
Quoted 8 lines · On Fri, Jun 26, 2026 at 2:40 AM <[email protected]> wrote:

From: luoxuanqiang <[email protected]>

IP_LOCAL_PORT_RANGE lets applications override the netns ephemeral port range on a per-socket basis. __inet_hash_connect() already treats such a range as an explicit application partition and scans it with step 1 [1].

Do the same in inet_csk_find_open_port():

What's the use case of IP_LOCAL_PORT_RANGE + bind(, 0) without IP_BIND_ADDRESS_NO_PORT ?

Quoted 76 lines

when a socket-local range is set,
walk the whole selected range instead of first splitting it by parity. Keep the existing step-2 parity behavior for sockets using the netns range, so the default bind/connect separation remains unchanged.

[1] https://lore.kernel.org/r/[email protected]

---
 net/ipv4/inet_connection_sock.c | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)
net/ipv4/inet_connection_sock.c +13 -8
index 56902bba54838..ad8af70c92ca3 100644
@@ -323,13 +323,16 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
        struct inet_bind2_bucket *tb2;
        struct inet_bind_bucket *tb;
        u32 remaining, offset;
+       bool local_ports;
        bool relax = false;
+       int step;
 
        l3mdev = inet_sk_bound_l3mdev(sk);
 ports_exhausted:
        attempt_half = (sk->sk_reuse == SK_CAN_REUSE) ? 1 : 0;
 other_half_scan:
-       inet_sk_get_local_port_range(sk, &low, &high);
+       local_ports = inet_sk_get_local_port_range(sk, &low, &high);
+       step = local_ports ? 1 : 2;
        high++; /* [32768, 60999] -> [32768, 61000[ */
        if (high - low < 4)
                attempt_half = 0;
@@ -342,18 +345,19 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
                        low = half;
        }
        remaining = high - low;
-       if (likely(remaining > 1))
+       if (!local_ports && remaining > 1)
                remaining &= ~1U;
 
        offset = get_random_u32_below(remaining);
        /* __inet_hash_connect() favors ports having @low parity
         * We do the opposite to not pollute connect() users.
         */
-       offset |= 1U;
+       if (!local_ports)
+               offset |= 1U;
 
 other_parity_scan:
        port = low + offset;
-       for (i = 0; i < remaining; i += 2, port += 2) {
+       for (i = 0; i < remaining; i += step, port += step) {
                if (unlikely(port >= high))
                        port -= remaining;
                if (inet_is_local_reserved_port(net, port))
@@ -384,9 +388,11 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,
                cond_resched();
        }
 
-       offset--;
-       if (!(offset & 1))
-               goto other_parity_scan;
+       if (!local_ports) {
+               offset--;
+               if (!(offset & 1))
+                       goto other_parity_scan;
+       }
 
        if (attempt_half == 1) {
                /* OK we now try the upper half of the range */
--

2.43.0

#2 lore raw
Quoted 8 lines

2026年6月27日 07:40,Kuniyuki Iwashima <[email protected]> 写道:

Quoted 3 lines · On Fri, Jun 26, 2026 at 2:40 AM <[email protected]> wrote:

From: luoxuanqiang <[email protected]>

Quoted 8 lines
Quoted 5 lines

IP_LOCAL_PORT_RANGE lets applications override the netns ephemeral port range on a per-socket basis. __inet_hash_connect() already treats such a range as an explicit application partition and scans it with step 1 [1].

Do the same in inet_csk_find_open_port():

What's the use case of IP_LOCAL_PORT_RANGE + bind(, 0) without IP_BIND_ADDRESS_NO_PORT ?

Hi Kuniyuki,

Thanks for the question!

The use case is when an application wants to restrict ephemeral port allocation to a socket-local IP_LOCAL_PORT_RANGE, but still needs bind(..., 0) to allocate and reserve a local port immediately.

IP_BIND_ADDRESS_NO_PORT is useful when the application can defer port allocation until connect(), but it changes this behavior: bind(..., 0) does not reserve a port in that case. So it is not a replacement for applications that need the local port before connect(), for example to publish it to another component or set up local policy.

This patch is also intended to keep the bind(..., 0) path consistent with Eric's earlier change in __inet_hash_connect().

Thanks,
Xuanqiang

#3 lore raw
Quoted 14 lines · luoxuanqiang
Quoted 6 lines

2026年6月27日 07:40,Kuniyuki Iwashima <[email protected]> 写道:

Quoted 3 lines · On Fri, Jun 26, 2026 at 2:40 AM <[email protected]> wrote:

From: luoxuanqiang <[email protected]>

Quoted 8 lines
Quoted 5 lines

IP_LOCAL_PORT_RANGE lets applications override the netns ephemeral port range on a per-socket basis. __inet_hash_connect() already treats such a range as an explicit application partition and scans it with step 1 [1].

Do the same in inet_csk_find_open_port():

What's the use case of IP_LOCAL_PORT_RANGE + bind(, 0) without IP_BIND_ADDRESS_NO_PORT ?

Quoted 7 lines

Hi Kuniyuki,

Thanks for the question!

The use case is when an application wants to restrict ephemeral port allocation to a socket-local IP_LOCAL_PORT_RANGE, but still needs bind(..., 0) to allocate and reserve a local port immediately.

IP_LOCAL_PORT_RANGE was introduced for connect().

Unlike connect(), bind() occupies the port without SO_REUSEADDR/PORT, so I don't think the step 1 or 2 makes any difference.

Quoted 12 lines

IP_BIND_ADDRESS_NO_PORT is useful when the application can defer port allocation until connect(), but it changes this behavior: bind(..., 0) does not reserve a port in that case. So it is not a replacement for applications that need the local port before connect(), for example to publish it to another component or set up local policy.

This patch is also intended to keep the bind(..., 0) path consistent with Eric's earlier change in __inet_hash_connect().

Thanks,
Xuanqiang

#4 lore raw
Quoted 17 lines · Kuniyuki Iwashima
Quoted 16 lines · luoxuanqiang
Quoted 5 lines

2026年6月27日 07:40,Kuniyuki Iwashima <[email protected]> 写道:

Quoted 2 lines · On Fri, Jun 26, 2026 at 2:40 AM <[email protected]> wrote:

From: luoxuanqiang <[email protected]>

Quoted 7 lines
Quoted 5 lines

IP_LOCAL_PORT_RANGE lets applications override the netns ephemeral port range on a per-socket basis. __inet_hash_connect() already treats such a range as an explicit application partition and scans it with step 1 [1].

Do the same in inet_csk_find_open_port():

What's the use case of IP_LOCAL_PORT_RANGE + bind(, 0) without IP_BIND_ADDRESS_NO_PORT ?

Hi Kuniyuki,

Thanks for the question!

Quoted 8 lines
Quoted 3 lines

The use case is when an application wants to restrict ephemeral port allocation to a socket-local IP_LOCAL_PORT_RANGE, but still needs bind(..., 0) to allocate and reserve a local port immediately.

IP_LOCAL_PORT_RANGE was introduced for connect().

Unlike connect(), bind() occupies the port without SO_REUSEADDR/PORT, so I don't think the step 1 or 2 makes any difference.

Hi Kuniyuki,

That's a fair point — bind() takes exclusive ownership of the port without SO_REUSEADDR/PORT, so the parity split only changes the scan order, not the set of ports bind() can pick.  Correctness-wise there is no difference between step 1 and step 2 here.

There are a couple of smaller things that made me think it is still worth aligning the two paths, though:

- inet_csk_find_open_port() already consumes the narrowed range from IP_LOCAL_PORT_RANGE since commit 91d0b78c5177f ("inet: Add IP_LOCAL_PORT_RANGE socket option"), so the bind path isn't insulated from the option.  It just didn't pick up the relaxed scan step that __inet_hash_connect() got in commit 207184853dbdb ("tcp/dccp: change source port selection at connect() time").  Eric even noted at the end of that commit:

    "A similar change can be done in inet_csk_find_open_port() if needed."

  So this felt more like completing the companion change than adding something new.

- The entropy argument from commit 207184853dbdb applies equally here:
  the parity split drops one bit of the 16-bit sport for RSS hashing. Whether the port came from connect() or bind() doesn't matter to the NIC, and losing that bit hurts more when the application has already shrunk its port space with IP_LOCAL_PORT_RANGE.

- When IP_BIND_ADDRESS_NO_PORT is not set, plain bind(, 0) reserves a port immediately through inet_csk_find_open_port().  If the same application also uses connect() on other sockets within the same IP_LOCAL_PORT_RANGE, connect() now scans the full range while bind() still biases toward odd ports — so the parity heuristic works against itself inside the application's own partition.

Does that make sense, or am I over-thinking the consistency angle?

Thanks,
Xuanqiang

Quoted 11 lines
Quoted 11 lines

IP_BIND_ADDRESS_NO_PORT is useful when the application can defer port allocation until connect(), but it changes this behavior: bind(..., 0) does not reserve a port in that case. So it is not a replacement for applications that need the local port before connect(), for example to publish it to another component or set up local policy.

This patch is also intended to keep the bind(..., 0) path consistent with Eric's earlier change in __inet_hash_connect().

Thanks,
Xuanqiang